Gemma-4-26B-A4B-it-NVFP4A16

First community NVFP4 quantization of google/gemma-4-26B-A4B-it — the Mixture-of-Experts variant of Gemma 4 with 25.2B total parameters and only 3.8B active per token.

W4A16 — weights in FP4, activations in FP16 (weight-only quantization).

Key Specs

Serving with vLLM

Requirements

• vLLM build with transformers >= 5.4 (for Gemma 4 architecture support)
• On DGX Spark / SM 12.1: spark-vllm-docker built with --tf5 flag
• Included gemma4_patched.py for NVFP4 MoE scale key loading (see vLLM Patch)

Quick Start

docker run -d \
  --name vllm-gemma-4 \
  --gpus all --ipc=host --network host \
  -e VLLM_NVFP4_GEMM_BACKEND=marlin \
  -v /path/to/Gemma-4-26B-A4B-it-NVFP4A16:/model \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -v /path/to/gemma4_patched.py:/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/gemma4.py \
  <your-vllm-image> \
  vllm serve /model \
    --served-model-name gemma-4 \
    --host 0.0.0.0 --port 8888 \
    --quantization modelopt \
    --dtype auto \
    --kv-cache-dtype fp8 \
    --gpu-memory-utilization 0.85 \
    --max-model-len 65536 \
    --max-num-seqs 4 \
    --moe-backend marlin \
    --trust-remote-code

Key Flags

Testing

This is an instruct model — use the chat completions endpoint:

curl http://localhost:8888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4",
    "messages": [{"role": "user", "content": "Hello! Tell me a joke."}],
    "max_tokens": 200
  }'

DGX Spark

Tested on NVIDIA DGX Spark (128GB unified memory, GB10 Blackwell, SM 12.1). Model loads at 15.7 GiB — plenty of headroom for 65K+ context with FP8 KV cache.

How this was made

The Problem

Gemma 4 MoE stores expert weights as fused 3D tensors (nn.Parameter of shape [128, dim, dim]) instead of individual nn.Linear modules. NVIDIA Model Optimizer (modelopt) only quantizes nn.Linear — it silently skips the 3D expert parameters, which are 91% of the model.

The Solution

We wrote a _QuantGemma4TextExperts modelopt plugin that unfuses the 3D expert tensors into 128 × 3 individual nn.Linear layers before quantization. This follows the same pattern modelopt uses for Qwen3.5, Llama4, and DBRX MoE models. After quantization, a post-processing step renames the exported keys to match vLLM's expected format.

Calibration

• Tool: NVIDIA Model Optimizer v0.43, _nvfp4_selective_quant_cfg(["*"], weight_only=True)
• Data: 4096 samples from CNN/DailyMail, batch 16, seq_len 1024
• Expert routing: Natural (router decides which experts see which data)
• Vision encoder: Excluded from quantization (stays BF16)
• Hardware: NVIDIA DGX Spark

vLLM Patch

vLLM's Gemma 4 expert_params_mapping doesn't correctly map NVFP4 scale keys (.weight_scale, .weight_scale_2, .input_scale) to FusedMoE parameter names. The included gemma4_patched.py fixes this. A PR to upstream vLLM is forthcoming.

Reproduce

pip install torch transformers>=5.4 accelerate datasets
git clone https://github.com/NVIDIA/Model-Optimizer.git
pip install -e Model-Optimizer[all]
pip install --force-reinstall transformers>=5.4 huggingface_hub>=1.5

python quantize_gemma4_moe.py --qformat nvfp4_w4a16

Full quantization script included as quantize_gemma4_moe.py.

Limitations

• Requires vLLM with transformers >= 5.4 and the included gemma4_patched.py
• --moe-backend marlin required for correct MoE computation
• Community quantization, not an official NVIDIA or Google release

License

Apache 2.0 — inherited from the base model.

Credits

Quantized by Mario Iseli on an NVIDIA DGX Spark. Built and validated with AI-engineering assistance from Anthropic.

Shout-out to eugr/spark-vllm-docker for the DGX Spark-optimized vLLM build.

📬 mario@marioiseli.com ☕ Buy me a coffee if this makes your Spark go brrrrrr! 🚀